How can you tell if a picture of a person is AI-generated?
By FlirtAlly · October 4, 2026
How can you tell if a picture of a person is AI-generated?
Key Facts
- Humans spot AI-generated content just 51.2% of the time — statistically a coin toss, per a study of 1,276 participants.
- AI-generated human faces are the hardest to detect, with people identifying them correctly only 46.6% of the time, researchers found.
- A meta-analysis of 86,155 participants confirms image detection accuracy sits at 53.16% — indistinguishable from chance, the analysis shows.
- Prior familiarity with AI media offers zero advantage: experts scored 51.9% versus newcomers' 51.1%, the study revealed.
- Five leading AI detection tools disagreed with each other on 35 of 45 test images — nearly 8 in 10, a NewsGuard audit found.
- Detection models that ace lab benchmarks 'catastrophically fail' on real social media content after compression, academic research shows.
- Structured support like feedback training and AI assistance lifts detection accuracy to 65.14%, according to the meta-analysis.
Why Your Eyes Can't Be Trusted Anymore
Flip a coin. That's essentially your accuracy rate at spotting an AI-generated image in 2025 — and if the picture is of a human face, you're actually doing worse than the coin.
A preregistered study of 1,276 participants found people identify synthetic content at just 51.2% accuracy. Synthetic faces were the single hardest category, at 46.6% — worse than animals, food, or landscapes. The researchers' conclusion is blunt: relying on human perception "is no longer a viable bulwark against the threats posed by synthetic media."
For a while, spotting AI images was almost fun. Count the fingers. Check the teeth. Look for garbled text on signs. When fake "Trump arrest" photos circulated in 2023, malformed hands helped expose them — but that method no longer works, as reporter guidance from GIJN makes clear. Today's generators render anatomically correct hands, produce accurate text, and handle reflections and shadows convincingly.
The uncomfortable result is an asymmetry in perception: people correctly confirm authentic content 64.6% of the time but spot synthetic content only 38.8% of the time. We default to believing what looks plausible — which is precisely what modern generators optimize for.
Here's the finding that surprises even media professionals: prior familiarity with synthetic media provides no measurable advantage. The same study found detection accuracy essentially identical across familiarity levels — 51.1% for newcomers, 51.9% for experts. A separate meta-analysis of 56 studies covering 86,155 participants confirms the pattern, with image detection sitting at 53.16%, statistically indistinguishable from chance.
What does help, modestly, is structured support — feedback training, AI assistance, and caricaturized examples pushed accuracy to 65.14%. Still far from reliable, but it points to a principle: process beats perception.
If you can't trust your eyes, you need signals that don't depend on them. The checks that still carry weight are the ones grounded in provenance and consistency over time:
- Does the person look identically the same across many images, or does the face drift?
- Can you trace where the image came from — who published it, and when?
- Does the visual context hold up under geometric tests like shadow direction and vanishing points?
- Is the source transparent about whether the person is real or synthetic?
That last point is why platforms are increasingly designing for verifiability rather than hoping users can spot fakes. FlirtAlly, for example, locks a canonical face for each of its AI companions and identity-checks every generated photo and video against it — so the question shifts from "does this look fake?" to "is this consistently, verifiably the same person?" It's the same logic the research supports: identity consistency is a stronger signal than artifact-hunting, because the artifacts have largely disappeared.
As GIJN's guide warns, misplaced certainty is more dangerous than honest uncertainty. The honest answer to "can you spot an AI face?" is: probably not — and the sooner we stop pretending otherwise, the better our detection processes become.
Why Detection Tools Alone Won't Save You Either
If you're tempted to run a suspicious photo through an online detector and call it a day, the data says don't. A NewsGuard audit of five leading tools found they disagreed with each other on 35 of 45 test images — nearly eight in ten. Worse, false positives are rampant: one tool flagged a genuine proof-of-life video of Israeli Prime Minister Netanyahu as 96.9% likely AI-generated, a rating bad actors then weaponized to spread disinformation.
The problem isn't just vendor exaggeration. Academic research shows detection models that ace controlled benchmarks "catastrophically fail" on real-world content once compression, resizing, and platform processing erase the high-frequency traces they rely on. Learned features turn out to be overly specific to the artifacts of particular generators, not generalizable signals. In practice, that means a binary "AI vs. human" verdict from any single tool is effectively a guess.
This is exactly why FlirtAlly's media pipeline doesn't lean on a single detector. Every selfie, video, and couple photo a companion generates is identity-checked against her canonical locked face — a provenance signal that survives compression because it's built into the render, not hunted for after the fact.
A reliable workflow treats tool output as one input among many, not the answer:
- Visual anomaly checks — geometry, shadows, skin texture — that survive compression
- Provenance and consistency over time: does this image come from a known, verifiable source?
- Reverse image search and metadata analysis for context
- Human judgment informed by structured verification steps, not gut feeling
